[13a-S101-1~10] FS.1 Focused Session "AI Electronics"

[13a-S101-8] Robustness Evaluation of Direct Feedback Alignment Learning for Analog VLSI Neural Networks

〇(B)Naoki Nagamatsu¹, Hideki Murakami¹, Seiji Uenohara², Takashi Morie² (1.NIT, Kurume College, 2.Kyushu Inst. Tech.)

Keywords:Direct feedback alignment, Neural network, Analog integrated circuit

We focused on the direct feedback alignment (DFA) method as a learning algorithm for analog VLSI implementation of neural networks, and evaluated its robustness against non-ideal characteristics, which is a concern for analog VLSI implementation. It was found that DFA is more robust against the variation of device characteristics than the back propagation (BP) algorithm, and is suitable for implementations with analog memory devices.